Usage of Isotropic MRI Images Impoves Prostate Cancer Localization Results

MRI images are often anisotropic which make two dimensional neural networks perform better than three dimensional ones due to the fact they cannot take spacing between adjacent slices into account. The aim of this study was to validate if an earlier proposed technique for converting anisotropic images to isotropic enable three dimensional networks to yield better results than two dimensional networks and increase overall accuracy in the task of segmenting prostate cancer as it was shown to do in the task of segmenting prostate. In order to achieve the aim a combination of previously proposed image conversion technique as well as no-new U-Net (nnU-Net) was used to localize prostate cancer. Majority of published studies on prostate cancer segmentation deal with images acquired at a single institution, while this study deals with the image dataset gathered from 11 different institutions for both model training and validation allowing the assessment of model generalizability. The results of performed experiments confirmed that moving away from anisotropic images and two-dimensional neural network to isotropic images and three-dimensional neural network can improve accuracy while segmenting prostate cancer in MRI images. This study showed that nnU-Net is able to provide similar accuracy to other studies as well as the ability to distinguish clinically significant cancer from clinically insignificant one. This study also showed that a lot of identified prostate zones cannot neither be confirmed nor rejected as being abnormal due to the nature of how ground truth is established thus revealing the need of prospective accuracy evaluation.